Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of...
-
date post
21-Dec-2015 -
Category
Documents
-
view
219 -
download
5
Transcript of Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of...
![Page 1: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/1.jpg)
Model Choice in Time Series Studies of Air Pollution and Health
Roger D. Peng, PhDDepartment of Biostatistics
Johns Hopkins Blomberg School of Public Health
APHA 2007
Sponsors: NIEHS, EPA, Health Effects Institute
![Page 2: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/2.jpg)
City-specific estimates, PM10 and mortality, 1987-2000
![Page 3: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/3.jpg)
PM10, PM2.5 and Mortality, NMMAPS, 100 Cities
Dominici, et al 2007
![Page 4: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/4.jpg)
PM10 and Mortality: Sensitivity of the National Average Estimate to Adjustment for Weather (Welty, et al 2005)
![Page 5: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/5.jpg)
Which Estimate?
• In time series studies of air pollution and health, model choice is important
• Risk is difficult to estimate, signal-to-noise is weak
• Models are complex and risk estimates can be sensitive
• Risk estimates have substantial policy impact
![Page 6: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/6.jpg)
Some air pollution and mortality dataMortality and PM10 in San Francisco, 1998--2000
Mortality and PM10 in New York City, 1998--2000
![Page 7: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/7.jpg)
Potential Confounders in Air Pollution and Health Studies
• Smoothly varying seasonal trends• Long-term trends
– structural changes in overall population
• Temperature– mortality: “J-shaped” relationship– PM10: increasing/positive
• Humidity• Other time-varying factors?
![Page 8: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/8.jpg)
Confounding by Season: New York City
![Page 9: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/9.jpg)
![Page 10: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/10.jpg)
![Page 11: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/11.jpg)
![Page 12: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/12.jpg)
![Page 13: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/13.jpg)
Season-specific associations are positive, overall association is negative
![Page 14: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/14.jpg)
Stage 1: City-specific model
Poisson regression model
Pollutant series WeatherSeasonal, long-
term trends
![Page 15: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/15.jpg)
Simulation Study Model of Mortality and Air Pollution
• Mortality
yt ~ xt + f(t) + t
• Air pollution
xt ~ g(t) + t
What is the relationship between f and g?Theory indicates f should have
smoothness required to best predict xt
![Page 16: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/16.jpg)
Scenario 1
![Page 17: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/17.jpg)
Scenario 2
![Page 18: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/18.jpg)
Comparison of model formulations in the literature
• Representing s()– smoothing splines– penalized splines – natural splines
• Choosing df– minimize AIC (best predict mortality)– minimize residual autocorrelation (via PACF)– best predict pollution (GCV-PM10)
![Page 19: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/19.jpg)
Results of simulation study
• Bias drops off as df increases• Variance increases a little with df• Nonparametric smoothers require more df to
remove bias• AIC and PACF methods more biased in
Scenario 2
• Predicting xt via GCV had best performance overall (w.r.t. mean squared error)
![Page 20: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/20.jpg)
PM10 and Mortality: Sensitivity of National Average Estimates to Adjustment for Seasonal Trends
Peng, et al 2006
![Page 21: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/21.jpg)
National Ambient Air Quality Standards: Statistical research has an impact
• From US EPA NAAQS Criteria Document 1996: “Many of the time-series epidemiology studies looking for associations between ozone exposure and daily human mortality have been difficult to interpret because of methodological or statistical weaknesses, including the a failure to account for other pollutant and environmental effects.”
• From US EPA Criteria Document 2006: “While uncertainties remain in some areas, it can be concluded that robust associations have been identified between various measures of daily ozone concentrations and increased risk of mortality.”
![Page 22: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/22.jpg)
What Next?
• What are the mechanisms of PM toxicity?– size, chemical component, source
• Data for ~60 chemical components are available but are sparse in time and space
• Individual constituents can be highly correlated with each other
• Interactions are potentially important and difficult to identify
• Sources are not measured• Confounding issues are more complex
![Page 23: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/23.jpg)
PM2.5 PM10
PM10-2.5
Source-markersChemical constituents
Size Total mass
Cl
OC
SO4
Si
K
EC
NO3
Ca
Al
Fe
BiomassBurning
Vehicles
Crustal
Source: Bell et al EHP 2007
![Page 24: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/24.jpg)
What Next?
• In an increasingly complex setting, we need to be able to see the evidence in the data– Avoid lumping results together and providing “the
answer” (although sometimes we need a number)
• Reproduce published findings• Allow others to examine the data, check
models and test assumptions• Enable development of new models/methods• We need reproducible research
![Page 25: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/25.jpg)
Tools for Distributing Reproducible Research
• Doing research is primary, distributing it is often considered secondary– Is publishing and article enough?
• We lack an infrastructure for easily distributing statistical analyses to others (we often start from scratch each time)
• Need to automate the distribution process• Allow people to work in their natural
environment
![Page 26: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/26.jpg)
Summary
• Time series models of air pollution and health are potentially confounded seasonal trends
• Risk estimates can vary over a wide range depending on the model chosen
• National average estimates from multi-site studies are generally robust to model choices
• Reproducible research is needed to allow others to assess the evidence and provide transparency
![Page 27: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/27.jpg)
Collaborators
• Francesca Dominici• Tom Louis• Aidan McDermott• Luu Pham• Ron White• Jonathan Samet• Scott Zeger
• Michelle Bell (Yale)• Keita Ebisu (Yale)• Leah Welty (NWU)
![Page 28: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/28.jpg)
![Page 29: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/29.jpg)
Stage 1: City-specific model
Poisson regression model
![Page 30: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/30.jpg)
Stage 1: City-specific model
Poisson regression modelPollutant series
![Page 31: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/31.jpg)
Stage 1: City-specific model
Poisson regression model
Weather
![Page 32: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/32.jpg)
Stage 1: City-specific model
Poisson regression model
Seasonal and long-term trends
![Page 33: Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.](https://reader035.fdocuments.net/reader035/viewer/2022062714/56649d5f5503460f94a3ef7e/html5/thumbnails/33.jpg)
Designing Tools for theResearch Pipeline